From 192593ba07323244b99731a8a0aeddeb925735b6 Mon Sep 17 00:00:00 2001 From: Lj Miranda Date: Tue, 18 Jul 2023 09:34:16 +0800 Subject: [PATCH] Fix copied scores Accidentally copied the POS tagging scores, not the UAS --- projects/_posts/2023-08-07-calamancy.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/projects/_posts/2023-08-07-calamancy.md b/projects/_posts/2023-08-07-calamancy.md index f46c76ec..7d7afcde 100644 --- a/projects/_posts/2023-08-07-calamancy.md +++ b/projects/_posts/2023-08-07-calamancy.md @@ -145,9 +145,9 @@ We also evaluated cross-lingual and multilingual approaches in our benchmarks: | Language Pipeline | Binary textcat (Hatespeech) | Multilabel textcat (Dengue) | NER (TLUnified-NER) | Dependency parsing, UAS (Merged UD) | Dependency parsing, LAS (Merged UD) | |------------------------|---------------------------------------------------------|----------------------------------------------------------|-----------------------------------------------------|-------------------------------------|-------------------------------------| -| uk_core_news_trf | $$75.24 (0.05)$$ | $$65.57 (0.01)$$ | $$51.11 (0.02)$$ | $$54.77$$ | $$82.86$$ | -| ro_core_news_lg | $$69.01 (0.01)$$ | $$59.10 (0.01)$$ | $$02.01 (0.00)$$ | $$84.65$$ | $$82.80$$ | -| ca_core_news_trf | $$70.01 (0.02)$$ | $$59.42 (0.03)$$ | $$14.58 (0.02)$$ | $$91.17$$ | $$83.09$$ | +| uk_core_news_trf | $$75.24 (0.05)$$ | $$65.57 (0.01)$$ | $$51.11 (0.02)$$ | $$54.77$$ | $$37.68$$ | +| ro_core_news_lg | $$69.01 (0.01)$$ | $$59.10 (0.01)$$ | $$02.01 (0.00)$$ | $$84.65$$ | $$65.30$$ | +| ca_core_news_trf | $$70.01 (0.02)$$ | $$59.42 (0.03)$$ | $$14.58 (0.02)$$ | $$91.17$$ | $$79.30$$ | - **Multilingual**: we used XLM RoBERTa and an uncased version of mBERT as our base transformer models. We also finetuned each model for each task and did similar evaluations. Note that finetuning on XLM RoBERTa (both base and large versions) may require at least a V100 GPU. I've seen more consistent and stable training with an A100 GPU. Same can be said for mBERT.